Abstract

The water-energy-carbon nexus elucidates the potential for energy recovery and more sustainable solutions for energy-intensive water systems. Installing turbines or pumps as turbines (PaTs) in key areas of water distribution networks to recover energy can avoid wasted energy as well as reduce leakage caused by high water pressure. Previous studies have proposed models to predict PaT characteristics based on pump best efficiency point (BEP). These models apply statistical methods or artificial neural networks (ANNs) to determine the turbine characteristics. Nevertheless, other simpler machine learning methods could yield better predictive models. In addition to the PaT’s characteristics, its feasibility is also determined by the local operating hydraulic conditions. Previous PaT optimization studies generally consider one day of operation and few PaTs. The goal of the study is to develop a machine learning model to predict PaT performance, combined with a flexible optimization process to determine the most feasible selection and location of PaTs. For the predictive models, a database of 145 best efficiency points and 196 characteristic curves PaT experimental records were compiled. The machine learning models consistently outperformed all other models, including the current ANN, and previous models. A genetic algorithm (GA) was applied to maximize PaT power output and was applied to three example networks, modelled with EPANET. Combined with the large library of PaTs the GA was able to quickly converge to the best solution and find significant opportunities for energy recovery. Convergence on solutions for 3 distinct networks of different complexities was able to be determined within acceptable iterations. The solutions showed a significant power recovery of 44.16 MWh/year, 34.07MWh/year, and 124.22MWh/year with reasonable payback periods within 5 years.